Research On Dynamic Generalization Mechanism For Coreference Resolution

As the core of natural language processing (NLP), coreference resolution has the important significance in message understanding, text summarization, information extraction, information retrieval, information filtration, and machine translation. In this paper, we used instance-based dynamic generalization mechanism to solve coreference resolution in Chinese and English.The main idea of instance-based dynamic generalization mechanism is that we find all the instances similar with the testing instance from all training instances, with which we decide which class the testing instance belongs to. Based on such main idea, we proposed the concept of generalization point, and deviced two basic algorightm for dynamic generalization mechanism.In this thesis, we put emphasis on two kinds of dynamic generalization mechanism: Flat feature-based dynamic generalization mechanism and structured feature-based dynamic generalization mechanism.For research on flat feature-based dynamic generalization mechanism, we put emphasis on the problems of how to select best generalization point and define the confidence of be positive instance. We proposed five criterions on selecting best generalization point, and defined the confidence of be positive instance as piecewise linear function of the ratio of positive instance. The experimental result showed that flat feature-based dynamic generalization mechanism did as well as three traditional machine learning methods in Chinese and English corpus.Complexy feature is composed of character sequence-like feature and structured feature. In thesis, the research on complexy feature-based dynamic generalization mechanism is divided into two subtasks:(1) Research on head feature-based dynamic generalization mechanism. In thesis, we proposed the head of antecedent and anaphora as new features, which belong to character sequence-like feature. With error analysis of those two basic algorithm for dynamic generalization mechanism, we proposed competion mode for capturing the error of named entity recognition and semantic mutual exclusive collocation. The experimental result showed that with the competion mode, head feature-based dynamic generalization mechanism enhanced the resolution in English corpus, but it needs to be improved in Chinese corpus.(2) Research on structured feature-based dynamic generalization mechanism. In thesis, we proposed the Simple-Expansion tree-like structure as a new feature, which belong to structured feature. We proposed two pruning strategy for tree-like structure to solve the ploblem of structured generalization point match, and integrated the tree-like structured feature into dynamic generalization with competition mode. The experimental result showed that, with competion mode, structured feature-based dynamic generalization mechanism is undesirable for English corpus. The ressearch on exploiting the structured feature needs to be improved further.